Automated Assessment of Hypernasality in Motor Speech Disorders Using Nasal Cognate Distinctiveness and Nasal Posterior Probability

When a person speaks, the soft palate at the back of the mouth acts as a gate, directing air either through the nose or the mouth. In a healthy voice, this gate closes tightly for most sounds, opening only for specific letters like "m," "n," and "ng." However, for individuals with motor speech disorders—conditions where the brain has difficulty coordinating the muscles needed for speech—this gate may not close properly. The result is hypernasality, a condition where too much air escapes through the nose during sounds that should be purely oral, giving the voice a distinct, hollow quality. While speech-language pathologists can hear this difference, they often struggle to measure it precisely. Relying on the human ear alone can be inconsistent, making it difficult to track whether a patient is improving over time or to compare the severity of the condition across different people. To solve this, researchers have begun developing automated tools that listen to speech recordings and calculate exactly how much nasal resonance is present, turning a subjective impression into an objective number.

A team of researchers set out to test two specific automated methods designed to catch this nasal leakage. They wanted to know if these computer-based measures could accurately reflect the severity of hypernasality in patients with motor speech disorders. The first method, called Nasal Cognate Distinctiveness, focuses on pairs of sounds that are almost identical except for one being nasal and the other not. For example, the sound "p" is a voiceless stop, while "m" is its nasal partner. The computer analyzes how clearly a speaker distinguishes between these pairs; if the speaker has hypernasality, the "p" starts to sound too much like an "m," and the computer detects this blurring. The second method, known as nasal posterior probability, uses a more advanced computer model trained on the rules of language to estimate the likelihood that any given sound contains nasal energy. Unlike the first method, which traditionally looked only at stop consonants like "p," "t," and "k," this second approach was tested on a wider range of sounds, including other consonants and even vowels, to see if the nasal signal extends beyond the usual suspects.

To put these tools to the test, the researchers gathered sentence-level speech recordings from 561 people. This group included 374 individuals with typical neurological function to serve as a baseline, and 187 patients diagnosed with motor speech disorders. Each recording was also rated by trained speech-language pathologists who assigned a severity score based on their professional judgment. The researchers then ran the automated analyses on the recordings, looking for patterns. They found that as the human experts rated the hypernasality as more severe, the first measure, Nasal Cognate Distinctiveness, generally decreased, indicating that the distinction between oral and nasal sounds was becoming harder to detect. At the same time, the second measure, nasal posterior probability, increased, signaling a higher presence of nasal energy. The data showed that these changes were not random; they tracked closely with the human ratings of severity.

The study also investigated whether these automated measures could tell the difference between a speaker who had no hypernasality and one who did. When the researchers compared the performance of the two methods, they found that both were quite effective at making this distinction. The method focusing on sound pairs and the method analyzing the probability of nasal features both achieved similar success rates in separating the two groups. Perhaps more surprisingly, the researchers discovered that the second method worked just as well when applied to vowels as it did when applied to consonants. This finding challenges the older assumption that nasal problems are only visible in specific stop sounds. The data showed that the nasal signal is present and measurable across the entire spectrum of speech, from the sharp edges of consonants to the smooth flow of vowels, with both types of sounds showing the same strong relationship to the severity of the condition.

To ensure these tools were reliable enough for real-world use, the team also checked if the measurements would stay consistent if a person spoke the same sentence twice. They recorded 20 healthy controls on two separate occasions and compared the results. The analysis showed that for the vast majority of these repeat recordings, the measurements remained stable, with only a very small number of cases showing a change large enough to be considered a real difference rather than just normal variation. This step provided a crucial benchmark for understanding how much a score needs to change before a clinician can be confident that a patient's condition has actually improved or worsened. The study concludes that both automated approaches offer valid, objective ways to measure hypernasality. They capture the same underlying acoustic reality and provide a consistent, data-driven alternative to relying solely on the human ear, offering a clearer path for understanding and tracking motor speech disorders.